Wuhan Researchers Prove Compressed LLMs Hide Dangerous Uncertainty Behind Accurate Answers
A new benchmark shows that compressing LLMs can silently break their ability to know what they don't know, even when accuracy looks fine
Takeaways
- A new benchmark tests 12 LLMs (1B–70B) across quantization and pruning methods, finding that compression frequently breaks uncertainty estimation even when accuracy looks fine.
- Accuracy and uncertainty decouple under compression: a model can maintain top-1 accuracy while its prediction sets balloon, or lose accuracy while remaining overconfident.
- Larger models absorb compression-induced uncertainty ~6.9x better than smaller ones (Llama2-7B vs 70B under W4A4 QuaRot).
- Uncertainty inflation is threshold-like, not gradual — a model safe at 30% sparsity can collapse at 40–50%, making single-point evaluations dangerous.
- W4A16 (weight-only 4-bit) is the safest compression regime; W4A4 and structured pruning (especially SliceGPT) are the most destructive to uncertainty.
- The paper recommends sweeping compression ratios, measuring conformal prediction set size, and preferring larger base models when aggressive compression is needed.
When you quantize or prune a large language model to cut costs, you check whether accuracy holds up. That's the standard. But a new paper argues that check is fundamentally incomplete. A model can maintain its top-1 accuracy while becoming deeply confused about everything else , and in safety-critical settings, that hidden confusion is exactly what gets you into trouble.
The paper, from researchers at Wuhan University of Technology and Nanyang Technological University, presents the first systematic benchmark for evaluating how model compression affects uncertainty quantification in LLMs. They tested 12 models across four families (Llama-2, Llama-3, Qwen3, DeepSeek, Falcon) ranging from 1B to 70B parameters, under a wide range of compression configurations, using conformal prediction as their measurement tool.

Why accuracy alone isn't enough
The core problem is a phenomenon the authors call accuracy-uncertainty decoupling. Two models with the same accuracy can exhibit very different prediction set sizes, motivating uncertainty-aware evaluation beyond accuracy. Think of it like two students who both circle the right answer on a multiple-choice test , but one is certain, and the other is guessing among three plausible options. The score looks the same; the reliability is not.
In high-stakes applications such as medical diagnosis, legal reasoning, and autonomous decision-making, a model's ability to quantify what it does not know is as important as its predictive accuracy. Deploying a compressed model without understanding this gap introduces risks that can undermine the very efficiency gains compression was supposed to achieve.
The measuring stick: conformal prediction
To measure uncertainty rigorously, the authors use conformal prediction , a statistical framework that wraps any model's output into a prediction set rather than a single answer. Similar to the process of elimination used by students during exams, conformal prediction identifies a subset of potential labels in classification tasks by excluding improbable labels, which is statistically guaranteed to contain the true label, and quantifies uncertainty as the size of this subset. The coverage guarantee property makes conformal prediction a highly robust uncertainty quantification method.
The key metric here is